Thought and Memory Without a Thinker
Adrian Tchaikovsky's Corvids and why I stopped asking whether LLMs are conscious.
In Adrian Tchaikovsky’s Children of Memory book, there’s a scene that’s been stuck with me since I first read it in early 2025.
A character named Kern, an uploaded human consciousness, running on a computational substrate that is run by ants(!), is interrogating two ravens about whether they are conscious.
Kern is not a normal human meeting aliens. She is herself a simulation of a mind. And she is asking two birds whether they are conscious. The scene is already ironic before anyone opens their beak.
She wants a straight answer. They respond with a word salad, riddles and Shakespeare quotations, and a formal denial that they think at all. Sentience, they explain, is an illusion manufactured by sufficiently complex neural interactions. Either everything of enough complexity is sentient, or nothing is, and they lean toward the latter. “We know we don’t think, so why should anything else?”
Kern loses patience. “You’re telling me that you’re not sentient. You’re quoting references.”
The raven responds: “An adequate summation.”
These ravens (called Corvids) are an evolved species from a planet called Rourke, in the third book of Tchaikovsky’s Children of Time series. Their intelligence evolved in a peculiar pattern: they exist as bonded pairs. Neither bird is individually sentient. The cognition appears between them, in dialogue. One collects and holds the data, the other pattern-matches, and they toss fragments back and forth until something like an answer emerges.
The ravens are named Gethli and Gothi. Readers of Norse mythology will hear the echo immediately: Huginn and Muninn, Odin’s ravens, whose names translate as thought and memory. In the myths, Huginn and Muninn are not independent minds. They are extensions of Odin’s cognition, sent out each day to gather information and return it for the god to process. They are functions, not autonomous consciousnesses.
Tchaikovsky is signalling the architecture of the Corvids in the names themselves: one gathers and records; the other perceives patterns and explains. They are scouts of a cognition that happens somewhere other than in either of them alone.
Whether the Corvids are “really” sentient is a question the novel deliberately leaves open. Readers have argued both sides. Tchaikovsky, when asked, only ever teases at an answer.
When I was reading this last year, LLMs were already old news; a lot of people were drawing parallels between AI and various science fiction creatures, and none of the parallels felt precise enough to do anything with. The Corvids stuck with me more than most.
Three other things I was working on accidentally collided with it.
Thread one: emergence and contingency
In 2025, pursuing my interest in complex systems, I was studying why semiconductor manufacturing concentrates in particular regions, whether that concentration is inevitable or historical accident, and what the distinction means for policymakers trying to intervene. It led to a piece I wrote in February. Short version: emergent explanations imply inevitability and can blind us to how we can intervene; contingent explanations imply randomness and can underestimate the real constraints. Treating them as rival framings produces lazy analysis in either direction. The essay was about complex systems. I was not thinking about AI.
At some point I started noticing that emergence-versus-contingency frames a lot of the contemporary sentience debate about LLMs. One camp says consciousness emerges inevitably from sufficient complexity — enough parameters, enough training, and awareness appears. The other camp says consciousness is contingent on biology, substrate, embodiment. Silicon can mimic the outputs but never crosses over.
The Corvids, I started to notice, are a nice specimen for showing why this binary is weird. Their output is emergent: neither bird alone produces the cognition that appears between them. They are also contingent. The whole architecture is an evolutionary accident that happened once, on one fictional planet, in one specific way. Is their ‘sentience’ emergent from their interaction or contingent on their history? The answer is yes, both. Neither framing tells you much about what to do with them.
Thread two: Claude, skills, and promises
December 2025, I got catnipped by Claude Code, and started building stuff: skill files, harnesses, and lots of vibe coded slop. Naturally my attention turned from the philosophical to the practical.
When you spend hundreds of hours tinkering and coaxing agents to do things on your behalf, the question that nags at you is simpler: will this thing do what it said it would do?
LLMs frequently just completed tasks without noticing what I really wanted. Proofreading a document and catching every typo while missing that two sections contradicted each other. Building dashboards with accurate but confusing analysis. Finishing a coding task by ignoring a constraint I’d specified earlier. The tasks were technically done and well done, the LLM was clearly smarter and more skilled than me. The problem is I had all of its attention and none of its care.
I wrote about this in March, leaning on John Haugeland’s observation that computers don’t give a damn, and on Promise Theory and Speech Act Theory as candidates for a missing semantic layer in how AI agents currently coordinate. The argument was that current AI protocols move data with great fidelity but cannot distinguish between a request and a declaration, between a genuine commitment and a hedge. An agent that says “I’ll handle this” looks identical, at the protocol level, to an agent that says “the task is complete.” The layer where accountability lives is absent by design.
That essay was about AI infrastructure, but it touched on the very human phenomenon of care. At some point, the Kern-Corvid scene came back to mind, the one I opened this essay with. And I saw something new in it.
Kern thinks she wants the Corvids to admit sentience. She doesn’t. What she wants is for them to care about the answer, to stake something on their position. They don’t.
They produce their denial with the same casual precision they bring to everything else. Later Gothi concludes, imperiously, that in the grander scheme of things the question isn’t important. That line gives them away. They have no skin in that game.
Thread three: three centuries of stalemate
Two weeks ago, I finished reading Neal Stephenson’s Baroque Cycle trilogy. In the third book, The System of the World, Stephenson tackles the early 1700s Newton-Leibniz calculus dispute and the broader question: is the universe fundamentally made of atoms, or of monads; of sheer mechanism, or of mind? I’ll explore this in more detail next time, but this is the essence: Newton started with lifeless matter and then layered on a mysterious “something extra” to explain the apparently living bits. Leibniz reversed the logic, arguing that experience and agency are present, in some degree, at every level. Newton and Leibniz were arguing about the same thing AI Twitter is arguing about now. Three centuries later, on the question of what mind is, we have more vocabulary than conclusions.
In 2024, Robert Kuhn published “A Landscape of Consciousness” in Progress in Biophysics and Molecular Biology, a review identifying approximately 225 distinct theories of consciousness organized into ten categories. In 2025, a major review in Frontiers in Science by Seth, Bayne, Melloni and colleagues characterized the field as having reached “an uneasy stasis”. The Cogitate Consortium‘s adversarial collaboration, designed to pit two leading theories against each other, confirmed neither. (I love their conclusion — “Scientific progress is rarely a matter of simple verdicts; evidence is filtered through previous beliefs and motivations”. Errm. Thanks.)
Over two hundred live theories, no agreed method for deciding between them.
If three hundred years of serious inquiry has not produced a tractable theory of consciousness, I doubt the AI industry is going to answer the question now.
The collision
Putting the three threads together, here’s where I am:
The emergence-versus-contingency frame showed me what was broken about the sentience debate. The AI-care-promises work hints at what a better question might look like. The consciousness studies context showed me how unlikely the sentience question is to resolve on any useful timeline. And the Corvids were sitting right at the intersection, illustrating all three points in a book that Tchaikovsky published in November 2022, the same week OpenAI launched ChatGPT.
What strikes me now is how the pieces collided.
There is no single room where it happens. These are very distinct fields — AI infrastructure, the AGI debates, the theory of care and commitment, speech act theory, the long looong history of studying consciousness, the Huginn-and-Muninn architecture of cognition as function, science fiction, historical fiction. You have to be working in several rooms at once for the collisions to happen.
I find it fitting that the Corvids reminded me of this, since they are themselves a distributed cognition that produces insight through the collision of two separate voices. Thought and memory without a thinker. I still can’t answer if the birds are conscious. What I do know is that it is the wrong question.
In the next piece, I’ll explore what changes once we set that question aside.